How People Really Use AI — The Surprising Truth From Billions of Real Interactions
Everywhere we look, AI is sold as a productivity tool — write faster emails, generate code, summarise documents, and get more done.
But a massive new study from OpenRouter, based on more than 100 trillion tokens of real user interactions, shows that the truth is very different.
This is one of the most detailed looks ever into how people actually use AI models like ChatGPT, Claude, LLaMA, DeepSeek, and others.
And many of the findings challenge what the industry has been telling us for the past few years.
Roleplay Is the #1 Use Case — Not Productivity
This is the most unexpected insight.
More than 50% of all open-source model usage is for:
- Roleplay
- Interactive fiction
- Character conversations
- Game-like storytelling
People are using AI less like a work assistant and more like a creative or interactive companion.
Tech companies have been pushing “AI for office tasks,” but users are exploring:
- Characters
- Alternate worlds
- Emotional conversations
- Gaming scenarios
- Creative writing loops
Over 60% of all roleplay tokens come from structured, game-like interactions — not casual chat.
This is a completely different picture of AI usage than what’s publicly discussed.
Programming Has Become AI’s Fastest-Growing Use Case
While roleplay dominates open-source usage, coding-related queries have exploded across all models.
At the start of 2025:
- Programming made up just 11% of total usage.
By the end of 2025:
- Programming jumped to 50%+.
Developers are no longer asking for small code pieces. They are using AI for:
- Large debugging sessions
- Architecture reviews
- Multi-step problem solving
- Full codebase analysis
Average coding prompt size:
- Grew from 1,500 tokens → 6,000+ tokens
- Many exceed 20,000 tokens
Anthropic’s Claude leads this category with 60%+ share, though competitors are catching up.
The Rise of Chinese AI Models
Chinese AI has gone from a minor player to a global force.
Start of 2025:
- 13% share
End of 2025:
- ~30% of global usage
Leading models:
- DeepSeek
- Alibaba’s Qwen
- Moonshot AI
DeepSeek alone processed 14.37 trillion tokens.
Chinese models are growing fast because they are:
- Cheaper
- Very capable
- Popular in Asian markets
Languages:
- English still dominates at 83%
- Simplified Chinese is now 2nd at 5%
Asia’s share of AI spending jumped from 13% → 31%, with Singapore becoming the 2nd-largest country by usage.
The Shift to “Agentic” AI
Another major trend: AI is no longer just answering single questions.
Instead, it is being used as an autonomous agent that:
- Performs multi-step tasks
- Calls external tools
- Tracks context over long conversations
- Solves problems with planning
In early 2025:
- “Reasoning-optimised” tasks were almost zero.
End of 2025:
- More than 50% of interactions became agentic.
This means users now expect AI to act, not just respond.
Examples:
- “Debug this entire repository.”
- “Plan my product launch and create the assets.”
- “Analyse this dataset and generate insights.”
AI is becoming more like a co-worker than a search engine.
The “Glass Slipper Effect”: First Match = Long-term Loyalty
The study revealed a powerful behavioural effect.
When a model becomes the first to solve a user’s important problem:
- The user stays loyal
- Retention skyrockets
- Switching becomes unlikely
Example:
- Google Gemini 2.5 Pro’s June 2025 user cohort has 40% retention at month 5 — much higher than later cohorts.
This effect explains why some models build long-term communities very quickly.
Price Doesn’t Control Usage
Common belief:
Cheap models will win.
Reality from the study:
- A 10% price drop increases usage by only 0.5–0.7%
- Premium models still dominate because quality matters more
- Budget and premium models co-exist with strong usage
The LLM market isn’t a race to the bottom.
People choose AI models based on reliability, reasoning quality, and capability — not just price.
What This Means for the Future
The OpenRouter study reveals three big truths:
1. AI is not being used the way companies expected
People are using AI for creativity, companionship, coding, and long-form reasoning — not just typing emails.
2. The global AI map is changing
China now plays a major role, and Asia is growing fast.
3. AI is evolving into an agent
The next generation of AI will act more like a problem-solver and less like a text responder.
4. Early problem-solving wins loyalty
Whichever model solves a user’s biggest problem first becomes their long-term favourite.
Final Word
The gap between how we think AI is used and how it is actually used is much wider than expected.
Real-world data shows:
- People treat AI as a creative engine
- Developers rely heavily on AI for deep programming
- Chinese models are rising fast
- AI is moving toward autonomy
- Retention depends on meaningful problem-solving
Understanding these patterns will shape how businesses, creators, and learners adapt in the next phase of the AI revolution.
The post How People Really Use AI — The Surprising Truth From Billions of Real Interactions appeared first on FSIDM (Full Stack Institute of Digital Marketing).